GSM8k tightening submit_answer tool signature as zero-shot perf of submit_answer is 0% pass

[jamesbraza]

As of aviary.gsm8k==0.14.0, GSM8k's submit_answer tool accepts a str answer. When directly calling this tool, an LLM will tend to answer with an entire sentence for the str. This leads to 0% zero-shot performance of the submit_answer tool.

The way HotPotQA handles this is by using normalize_answer (string cleaning) then eval_answer with "contains" (in check) grading criteria.

We need to either move:

• GSM8k's submit_answer tool to accept a float answer
• GSM8k's submit_answer tool to match HotPotQA with string cleaning then a contains check